How effective is TabStructNet in capturing the structure of a table-image into an XML? : A reproducibility report

Vedhas Pandit

Abstract: In this submission, findings when attempting to reproduce the results from the EECV 2020 article [3] by using the corresponding code-repository github.com/sachinraja13/TabStructNet (made available by the authors of [3], i.e., by S. Raja et al.) are reported. Each challenge encountered, along with the corresponding solution -- which was either discovered or was learnt from the first author of [3] himself -- is described step-by-step. As a consequence, the intermediate files that one would manage to (and one needs to) generate at those steps, along with their inter-relationships with the rest of the code-repository have also been detailed. A few recommendations are put forward in process which might help the authors to make the repository more consistent with the paper, user-friendly, and as a consequence, to make the experiments more easily reproducible. In this submission, a few minor deviations of the model architecture from what is described in [3] to what is observed in the TabStructNet code repository are also reported. It is hoped that this report will make it easier for everyone to use and/or rebuild the described TabStructNet model.

Paper Url: https://openreview.net/forum?id=iB7gCxfLlMJ&noteId=taZ3YzgrjU-